The brand-fit dossier — Creator Overview, Deep Analysis, and Background Research

Why three tiers, not one big report

Brand managers don't read 5 paragraphs of LLM prose to decide whether a KOL is worth a second look. They read a one-line headline, glance at four star ratings, and either commit 30 seconds to the deeper read or move on. So the three tiers map onto the actual triage funnel:

1. Overview (1 credit, cached 30 days) — answers "what does this person do, and would I pursue them?" with a scored verdict. Reads 50 recent videos + author demographics + audience data when present.
2. Background Research (1 credit, cached 30 days) — answers "who is this person off-platform?". A search-LLM (openai/gpt-4o-mini-search-preview) browses public sources and returns Markdown with [1] citation chips. Useful for legitimacy / identity checks before a real outreach budget.
3. Deep Analysis (5 credits, cached 7 days) — for KOLs you've already shortlisted. Reads up to 200 videos + full audience snapshot + cross-platform contacts and produces five extra sections you can paste into a brief: content quality, partnership history, outreach strategy, negotiation notes, competitive context.

Scored ratings — only when we have ground truth

The four scores (Recommendation / Content fit / Audience fit / Risk) only render when the workspace brand profile has at least one structured field filled in. Without that anchor we couldn't score "fit" against anything — the LLM would be inventing a comparison.

Fill in any of product_category, target_audience_text, target_geographies, or exclusions_text on /workspaces/{slug}/settings and the scoring block lights up on every dossier. The LLM is told what to score against, asked to clamp to 1-5, and required to explain its reasoning in a 300-character score_rationale.

Every prose result lands in Postgres

Each of the three tiers has its own cache table — keyed by (workspace_id, username, platform), with the locale stashed inside the JSON payload so a re-render in a different language is a clean miss, not a stale read.

Operationally: a busy workspace opening the same KOL dossier 20 times in a week makes one LLM call to each tier. The 19 subsequent opens are free. The "Refresh" button on each card is the only path that burns a fresh call — a deliberate operator action, never an accidental fetch.

Markdown renderer, not whiteSpace: pre-line

Operator feedback was direct: "如果是 markdown，就用 markdown 格式 来优化展示". Every prose field is now rendered via a tiny inline Markdown component — headings, bold, italic, ordered + unordered lists, and the [1] / [2] citation chips for Background Research.

We didn't ship a full Markdown library; the renderer is single-file (~110 lines) and only handles the subset the LLMs actually emit. Smaller surface, no MDX plugin churn.

Locale, end to end

• NEXT_LOCALE cookie set by the locale switcher
• BFF reads the cookie and forwards ?locale=
• Railway service injects the locale into the LLM system prompt ("Respond entirely in Simplified Chinese (zh-CN)…")
• Cache key includes locale so en-cached bodies never serve zh callers

Three locales currently shipped: en / zh-CN / ja. Adding a fourth is a 3-line addition to _LOCALE_NAME in api/services/kol_overview.py plus the messages file.

Try the dossier

Open any KOL on /kols/[username]. If your workspace brand profile is set, you'll see star ratings on first load. Click Deep Analysis to see the five extra sections — billed against your workspace credit pool with explicit confirmation copy on the button.